Passage-Aware Structural Mapping for RGB-D Visual SLAM

Simultaneous Localization and Mapping (SLAM) has emerged as a fundamental capability of modern autonomous robots, allowing them to estimate their pose while incrementally reconstructing the surrounding environment. Among the available sensing modalities in SLAM, vision sensors provide a cost-effective means of capturing rich visual and structural data, leading to the emergence of Visual SLAM (VSLAM). Despite considerable progress in VSLAM, challenges remain in retrieving geometric information, especially in scenarios involving indoor navigation. Doors and passages, as critical structural elements in indoor environments, remain underexplored in existing VSLAM frameworks. By augmenting VSLAM with semantic information, more interpretable and structurally meaningful map reconstruction can be achieved, opening new possibilities for further research and applications.

In indoor environments, doors and passages are key elements that define rooms and establish their inter-connectivity. However, existing VSLAM methods fall short in detecting and modeling these elements. This is primarily due to the lack of effective detection of geometric discontinuities in walls and insufficient integration of semantic and topological information for doors and passages. Addressing this issue is crucial for improving the robustness and efficiency of indoor navigation, especially in complex and dynamic environments. Current methods often rely on prior environmental markers, limiting their practicality and scalability. Therefore, developing a method that can reliably detect doors and passages without relying on environmental markers is a significant challenge in current research.

The core innovations of this paper include the introduction of a passage-aware structural mapping method that detects doors and traversable openings by jointly fusing geometric, semantic, and topological cues. Specific innovations include: 1) Modeling doors as planar entities embedded within walls and classifying them as traversable or non-traversable based on their coplanarity with the supporting wall; 2) Proposing two complementary strategies for passage inference: traversal evidence accumulated from camera-wall interactions across consecutive keyframes, and geometric opening validation based on discontinuities in the mapped wall geometry; 3) Integrating this method into vS-Graphs as a proof of concept, enriching its scene graph with passage-level abstractions and improving room connectivity modeling. These innovations provide a novel solution for indoor robot navigation, addressing the underexplored area of door and passage detection in traditional Visual SLAM frameworks.

The methodology of this paper includes the following key steps:

• �� Input: Given an RGB-D point cloud at the VSLAM KeyFrame level, a panoptic segmentation method such as YOSO is used to extract semantically meaningful planar entities, including walls and doors.

• �� Process: Each KeyFrame is defined with pixel-wise semantic labels and instance-level masks, which are projected onto the point cloud to obtain semantically segmented point subsets.

• �� Output: RANSAC plane fitting is applied to estimate semantically validated planar entities, and detected entities are inserted into the map for continuous structural reconstruction.

• �� Passage Inference: Passages are inferred through two strategies: 1) Traversal evidence accumulated from camera-wall interactions; 2) Geometric opening validation based on discontinuities in the mapped wall geometry.

The experimental design includes data collection in indoor office environments using the SMapper device to validate the effectiveness of the proposed passage detection pipeline. The benchmarks used include vS-Graphs as the baseline, with evaluation metrics including the accuracy and robustness of doorway detection. Ablation studies are conducted to verify the effectiveness of the geometric opening validation strategy in reducing false positives. Key hyperparameters include threshold settings for RANSAC plane fitting and distance and angle thresholds in passage inference strategies. Comparative experiments demonstrate the performance advantages of the proposed method in complex indoor environments.

Experimental results demonstrate that the proposed method reliably detects doorways in indoor office sequences, laying the groundwork for future BIM-informed SLAM applications. Compared to traditional methods, this approach significantly enhances the accuracy and robustness of doorway detection, especially in complex indoor environments. Comparative experiments verify the effectiveness of the geometric opening validation strategy in reducing false positives, particularly in scenarios with occlusions or complex backgrounds. The experiments also show that the proposed method enriches the scene graph's passage-level abstractions while maintaining real-time performance, improving room connectivity modeling and scene understanding.

The application scenarios of this method include indoor robot navigation, Building Information Modeling (BIM), and smart building management. By detecting and modeling doors and passages, robots can better understand the structural environment, improving navigation and path planning efficiency. In BIM applications, this method can be used to verify and update building models, enhancing the intelligence of building management. The method can also be used in smart building management systems to improve building safety and energy efficiency by detecting and monitoring passage usage.

Despite the excellent performance of this method in detecting doors and passages, it may perform poorly in dynamic environments or with fast-moving cameras, as it relies on camera-wall interactions across consecutive keyframes. Additionally, in extreme lighting conditions, the performance of RGB-D sensors may affect the accuracy of door and passage detection. The method has primarily been validated in indoor environments and has yet to be tested in larger-scale or more complex environments. Future research directions include incorporating doors and passages directly into factor graph optimization for tighter coupling between traversability reasoning and pose estimation.